Vivaan Jain

Artha builds a point in time NSE panel from primary bhavcopy and declared corporate actions, delistings included, ranks the cross section on weekly momentum, constructs a 25 name book with Ledoit-Wolf minimum variance weights and Gârleanu-Pedersen partial adjustment, targets volatility, caps participation against ADV, and executes at the T+1 close. The full Indian cost stack is charged: STT, stamp duty, exchange and SEBI fees, GST, flat DP charges, and square root market impact.

The larger half of the project is proving the result is not luck. Validation is purged walk forward plus combinatorially purged cross validation, deflated Sharpe measured against an append only trial ledger, White's Reality Check and Hansen's SPA. The live layer holds backtest and engine parity as a CI gate, uses deterministic idempotent order ids, enforces drawdown rails, and runs a heartbeat that alarms on silence rather than only on failure. One line: a research lab that happens to trade, built so its own results cannot lie to it.

Synthetic total return benchmark because no free NIFTY 500 TRI exists; a static current sector map; cash earns 0%; hedge margin financing unmodelled; paper slippage degenerate until live quotes arrive; an announcement taxonomy audited at 81% accuracy; GDELT news coverage reaching 72 of 115 months before the free API rate limited it; and a deflated Sharpe of 0.20 that counts every single name experiment against the cross sectional book, which is deliberately conservative. Each one is tracked with a confirmation date, because a limitation you disclose cannot ambush you later.

Most backtesting frameworks leave the hard part to the user's discipline: not reading the future, not trusting a fill that would never happen, not forgetting how many strategies were tried before this one looked good. FullBacktester moves each of those from a habit into a construction constraint. A strategy never receives a full price history, only a PanelView of bars up to and including the current one, with read only arrays and bars stamped at their close instant in UTC, so a daily bar is not knowable until the exchange has actually closed.

Every strategy in an Arena runs against the same panel, the same execution config, and the same battery of checks, and the overfitting statistics use the number of entries as the number of trials. The summary reports CAGR, Sharpe, deflated Sharpe given that trial count, max drawdown, turnover, cost drag, and the Sharpe gap between the idealized and realistic engines, with every HIGH severity validation finding printed underneath.

A Transformer encoder built from scratch in TensorFlow that outputs directional forecasts for 4 horizons simultaneously: 1, 5, 10, and 20 trading days: from one forward pass. Trained on 4,900+ trading days (Jan 2006–Feb 2026). FinBERT news sentiment fused as a parallel signal. Achieved 0.62 AUC and 0.6+ non overlapping long short Sharpe on the held out test set.

A rigorous controlled experiment testing a specific hypothesis: does news sentiment provide incremental predictive signal beyond pure price based models in financial markets? Pipeline fuses historical price, volume, and API sourced news sentiment. Found a 5-day return correlation improvement from 0.43 to 0.46 using sentiment features, supported by MAE and RMSE diagnostics and shuffle testing. Conclusion: gains were modest, unstable, and highly sensitive to validation design: which is itself a meaningful finding about the fragile nature of financial signals.

Given the close on day T for roughly 208 NSE stocks, the pipeline predicts the overnight gap into the open on day T+1: how large the move will be, which direction it goes, and how confident the model is about each. These are four separate, independently fitted models. Magnitude is a LightGBM regressor on log1p(|return|), log space because the target is right skewed and a handful of genuinely extreme moves sit far out in the tail. Direction is a separate LightGBM classifier with its own features and its own hyperparameters, trained independently rather than derived from magnitude.

Direction confidence is deliberately not the classifier's raw probability, because tree ensemble probabilities are typically miscalibrated. Magnitude confidence is fitted on genuinely out of fold errors via a walk forward split on train, never on in sample residuals, because in sample residuals answer an easier question than the one that matters at prediction time.

Credit scoring sits at an intersection of three unsolved problems. First, high performing deep models learn complex non linear borrower representations that are impossible to explain to risk teams or satisfy regulators, explainability is now a legal requirement under frameworks like Basel and GDPR, and post hoc methods like SHAP produce approximations, not faithful explanations. Second, misclassification costs in lending are fundamentally asymmetric and instance dependent: rejecting a creditworthy borrower costs differently to approving a defaulter, and that ratio varies per loan. Optimising for AUC ignores this entirely. Third, existing models treat borrower risk as stationary, they miss that financial distress typically accelerates non linearly before a default event.

This research addresses all three with one unified architecture: learn latent risk embeddings, explicitly model debt acceleration dynamics, optimise directly on economic loss with an instance dependent cost function, and keep the decision process inherently interpretable.

A model that correctly identifies 90% of defaults but misclassifies every large loan defaulter as creditworthy scores well on AUC and is catastrophic in practice. This research trains directly on business loss: the asymmetric financial cost of each misclassification type, weighted by the specific borrower's loan profile. The evaluation metric is expected economic loss on the held out set, not AUC.

An end to end supervised learning pipeline over historical loan application data: exploratory analysis, preprocessing, feature engineering, and a binary classifier that predicts whether an application should be approved. A Flask interface wraps the trained model so applicant details can be entered and scored directly, which is the part that turns a notebook into something a person can actually use.

FinIntel is a multi agent system for equity research on India's National Stock Exchange. A fleet of specialised LLM agents, each grounded in live market data, lets a user analyse a single stock, screen the market by custom criteria, audit portfolio risk, or request a full research report. A five question suitability profile gates every recommendation against the user's stated risk tolerance, and human in the loop approval gates keep a person in control of consequential steps.

Every output passes a mandatory compliance pipeline: input validation, semantic screening, grounded rewrites that avoid fabrication, PII scrubbing, and SEBI style disclaimers. Deterministic mathematics (volatility, Sharpe, beta, drawdown, CAGR, XIRR, stress tests) is computed in Python; the LLM interprets the results rather than performing the arithmetic, which removes a whole class of hallucination.

A multi tenant RAG system where users upload documents and query them with natural language. Under the hood it uses three retrieval methods fused together: semantic vector search, BM25 keyword matching, and FlashRank cross encoder reranking combined via Reciprocal Rank Fusion. This eliminates the blind spots that any single retrieval method has. Real time streaming responses, multimodal document and image ingestion, shareable conversation links.

Built to run reliably on constrained 512MB infrastructure. Diagnosed and resolved 6+ critical production incidents including memory leaks, race conditions, and rate limit cascades. RAGAS evaluation suite covers faithfulness, context precision, and context recall across 25+ test cases.

Upload a resume, paste a job posting, and in 10 to 20 seconds ApplyPilot returns a cover letter, an application email, a fit score, adapted resume bullets, and grounded answers to custom application questions. Resume and job parsing extract structured profiles; the fit score is computed in plain Python before any model call; generation runs through Groq's free Llama 3.3 70B. A real status pipeline (ready → approved → submitted → interviewing → offered/rejected) tracks every application.

The AI never invents experience. Resume adaptation carries explicit prompt level constraints: reword existing bullets to hit JD keywords, never add a skill or change a date. It was tested specifically by feeding a resume missing an obvious JD keyword to confirm it reframes rather than fabricates.

Processes noisy field interview recordings in 13 Indian languages into clean, speaker labelled transcripts with zero paid API dependency. Raw audio goes through signal preprocessing, ASR via Whisper large-v3, speaker diarization via SpeechBrain ECAPA-TDNN, and semantic question mapping to produce structured outputs from unstructured conversational audio at scale.